schneewolflabs/A1
schneewolflabs/A1 is a 12 billion parameter, Mistral Nemo-class causal language model developed by schneewolflabs, specifically fine-tuned for explicit chain-of-thought reasoning. It leverages a Qwen3-style chat template to produce detailed reasoning traces within tags before generating its final answer. This model excels at complex problem-solving by explicitly showing its thought process, making it suitable for tasks requiring transparent and verifiable reasoning.
Loading preview...
Overview
schneewolflabs/A1 is a 12 billion parameter language model, derived from schneewolflabs/A0i-12B (Mistral Nemo-class), that has been supervised-fine-tuned to perform explicit chain-of-thought reasoning. Unlike its base model, A1 generates a reasoning trace within <think></think> tags, following the Qwen3 thinking convention, before providing a final answer. This capability was achieved by repurposing ten unused reserved token slots in the A0i tokenizer without resizing the vocabulary, initializing their embeddings, and training them during the fine-tuning process.
Key Capabilities
- Explicit Chain-of-Thought Reasoning: Generates detailed step-by-step reasoning within
<think></think>tags, enhancing transparency and verifiability of its answers. - Problem Solving: Demonstrates coherent reasoning that generalizes beyond its training distribution, capable of solving complex problems like the bat-and-ball problem correctly.
- Qwen3-style Chat Template: Utilizes a specific chat template to automatically inject reasoning prompts, ensuring the model engages its reasoning capabilities.
Training and Limitations
A1 was fine-tuned using LoRA on attention and MLP projections, along with embed_tokens and lm_head, over a single epoch on the schneewolflabs/BigDenker-SFT dataset. While it shows strong reasoning generalization, current limitations include an "always-on" thinking mode (even for trivial prompts) due to the template, and reasoning phrasing homogeneity from being trained on a single dataset. It is a text-only model and does not support vision or tool-calling, despite the presence of related tokens for chat-template completeness.